Importance Sampling-based Post-Processing Method for Global Sensitivity Analysis

Abstract

A methodology is presented for the calculation of global sensitivity indices from an existing sample of simulations. A sample weighting scheme similar to kernel regression is proposed for estimating the moments of the conditional distributions in the calculation of the sensitivity indices. The weights are developed based on importance sampling, but with the roles of the target distribution and the sampling distribution reversed. The method is demonstrated with benchmark examples from the sensitivity analysis literature.